Assessing the Predictive Utility of Logistic Regression, Classification and Regression Tree, Chi-Squared Automatic Interaction Detection, and Neural Network Models in Predicting Inmate Misconduct

Abstract

This study assesses the relative utility of a traditional regression approach - logistic regression (LR) - and three classification techniques - classification and regression tree (CART), chi-squared automatic interaction detection (CHAID), and multi-layer perceptron neural network (MLPNN)—in predicting inmate misconduct. The four models were tested using a sample of inmates held in state and federal prisons and predictors derived from the importation model on inmate adaptation. Multi-validation procedure and multiple evaluation indicators were used to evaluate and report the predictive accuracy. The overall accuracy of the four models varied between 0.60 and 0.66 with an overall AUC range of 0.60–0.70. The LR and MLPNN methods performed significantly better than the CART and CHAID techniques at identifying misbehaving inmates and the CHAID method outperformed the CART approach in classifying defied inmates. The MLPNN method performed significantly better than the LR technique in predicting inmate misconduct among the training samples.